NGARiAI/ngari-tool

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:1.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 4, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

NGARiAI/ngari-tool is a 1.5 billion parameter tool-calling model based on Qwen2.5-1.5B-Instruct, specifically fine-tuned for robust function calling. It achieves 100% tool detection, tool-name accuracy, and parameter validity on its internal evaluation, making it highly reliable for structured JSON output. This model is optimized for on-device agents, enabling fast, lightweight tool triggering and API call formatting without cloud dependency, suitable for edge deployments with 8GB RAM.

Loading preview...

NGARi Tool — Tool-Calling 1.5B

NGARiAI/ngari-tool is a specialized 1.5 billion parameter model designed for precise tool-calling and function invocation. Built upon the Qwen/Qwen2.5-1.5B-Instruct base model, it incorporates a LoRA adapter fine-tuned on synthetically generated multi-turn tool-calling scenarios. This unique teacher-student pipeline, where a 27B-class model generates perfect examples, enables ngari-tool to achieve exceptional accuracy in tool detection and parameter formatting.

Key Capabilities & Features

  • High Accuracy Tool Calling: Achieves 100% tool detection rate, 100% tool-name accuracy, and 100% parameter validity on the ngari-tool-stable_tool_eval.json benchmark.
  • Edge-Optimized Performance: Designed for on-device agents, it runs efficiently on hardware like the NVIDIA Jetson AGX Orin with 8GB RAM, facilitating fast, local execution without cloud reliance.
  • Structured Output: Excels at formatting function calls and structuring JSON outputs reliably, crucial for agentic workflows.
  • Apache 2.0 Licensed: Both the base model and NGARi-authored artifacts are available under the Apache 2.0 license.
  • Sovereign AI: Developed with a focus on user-owned edge hardware and air-gap verification, promoting complete local control.

Ideal Use Cases

  • On-device Agents: Perfect for deploying intelligent agents directly on edge devices that need to interact with local tools or APIs.
  • Offline Function Calling: Suitable for applications requiring robust tool invocation in environments with limited or no internet connectivity.
  • Resource-Constrained Environments: Its lightweight 1.5B parameter size and efficient design make it ideal for hardware with limited memory and processing power.
  • Reliable JSON Generation: When precise and valid JSON output for tool arguments is critical, this model offers high fidelity.

Finetunes
1 models